OCR gives you text. You needed answers.

The classic pipeline for getting data out of documents is OCR first: flatten every page into a stream of characters, then attack the stream with regexes, key-value heuristics, templates per form version, or lately a language model. We built that pipeline. It is why this product exists.

Where the text-first pipeline loses

• Structure dies in the flattening. A pricing table is meaning carried by geometry: which number sits under which column, which row it shares. Serialized to text, that geometry is gone, and no amount of downstream cleverness reliably reconstructs a merged-cell table from a character stream. This is the failure that killed our own first version.
• Checkboxes and marks are not text. Filled forms speak in ticks, circles, and strike-throughs. OCR either drops them or renders them as noise, and the difference between option B2 selected and not selected is the whole answer.
• Templates rot. Coordinate-based extraction breaks the day the form adds a line. Every vendor revision becomes maintenance, and every new document type is a project.
• No provenance. Once the page becomes a string, "where did this value come from" has no good answer, so a human who wants to verify a number re-reads the document, which was the job you were automating.

The schema-first alternative

Thresh skips the flattening. Vision-native models read the page as a page: layout, tables, marks, and handwriting together. You describe what you need (fields, tables with their columns, plain-language questions) and the result conforms to that schema. Every value carries the page it came from and a confidence flag; missing values come back explicitly as missing.

The difference is measurable, and we publish the measurement: 98.6% field accuracy and 97.2% table cell F1 across a 37-document corpus of real filled paperwork, including the dense pricing tables that break text-first pipelines. The methodology, corpus composition, and scoring rules are public.

When OCR is still the right tool

Honest scoping: if you need full-text search over an archive, a searchable PDF layer, or every word on the page rather than specific answers, OCR is the right technology and a mature one. Schema-first extraction earns its keep when the output is a decision or a system of record: specific fields, verified against the source, delivered as data.